Maximum kernel execution time for secondary card?

Accelerated Computing CUDA CUDA Programming and Performance
1

What is the maximum kernel execution time for secondary card in winXP?

I have a following problem with CUDA computations on GPU NVidia G92 family
( 8800 GT + 8800 GTS 512 , driver 177.85, MB 650i):

If experiments are run on the secondary graphics card (that did not run the system desktop), calculations that last for more than 10-11 seconds are killed.

The secondary graphics card is not associated with winXP desktop,
and I used the very simple kernels (e.g. BlackScholes_kernel.cuh from SDK
(only with big loop)).

2

It is a known bug in the driver, it will be fixed soon.

3

Thanks for answer.
Is it only for winXP and what about Linux?

4

As far as I know, Linux users does not suffer from it. Watchdog problem is Windows-specific.

5

It occurs in Linux if you are running X. Linux booted directly to console (you don’t run X beforehand) works fine.

6

In fact it is possible to boot in X mode if there is another graphics card to handle the display and the CUDA card is not attached to a X screen nor display device. Both cards need to be compatible with the unique driver. I use a 8400 for display and 8800gtx for computation, it works fine under X.

7

Yes, non-display cards don’t have any watchdog timer under X; I was referring to cards connected to monitors.

8

what would happen if you have say a pci or onboard ati graphics and one or two nvidia 8800s or 280s undex X? i assume it would work.

9

yes, assuming you run the script to initialize the device nodes for the NVIDIA cards.

10

mfatica Posted Sep 9 2008, 12:55 PM
It is a known bug in the driver, it will be fixed soon.

Thanks, this bug fixed now. Please see:
forum